Maximal Monotonicity, Conjugation and the Duality Product in Non-Reflexive Banach Spaces

Journal of Convex Analysis Volume 17 (2010), No. 2, 553 563 Maximal Monotonicity, Conjugation and the Duality Product in Non-Reflexive Banach Spaces M. Marques Alves IMPA, Estrada Dona Castorina 110, 22460-320 Rio de Janeiro, Brazil maicon@impa.br B. F. Svaiter IMPA, Estrada Dona Castorina 110, 22460-320 Rio de Janeiro, Brazil benar@impa.br Received: September 23, 2008 In this work we study some conditions which guarantee that a convex function represents a maximal monotone operator in non-reflexive Banach spaces. Keywords: Fitzpatrick function, maximal monotone operator, non-reflexive Banach spaces 2000 Mathematics Subject Classification: 47H05, 49J52, 47N10 1. Introduction Let X be a real Banach space and X its topological dual, both with norms denoted by. The duality product in X X will be denoted by: π : X X R, π(x,x ) := x,x = x (x). (1) A point to set operator T : X X is a relation on X X : T X X and T(x) = {x X (x,x ) T}. An operator T : X X is monotone if x y,x y 0, (x,x ),(y,y ) T and it is maximal monotone if it is monotone and maximal (with respect to the inclusion) in the family of monotone operators of X into X. The domain of T : X X is defined by D(T) := {x X T(x) }. Fitzpatrick proved constructively that maximal monotone operators are representable by convex functions. Before discussing his findings, let us establish some notation. We Partially supported by Brazilian CNPq scholarship 140525/2005-0. Partially supported by CNPq grants 300755/2005-8, 475647/2006-8 and by PRONEX-Optimization. ISSN 0944-6532 / $ 2.50 c Heldermann Verlag

554 M. Marques Alves, B. F. Svaiter / Maximal Monotonicity, Conjugation and... denote the set of extended-real valued functions on X by R X. The epigraph of f R X is defined by E(f) := {(x,µ) X R f(x) µ}. We say that f R X is lower semicontinuous (l.s.c. from now on) if E(f) is closed in the strong topology of X R. Let T : X X be maximal monotone. The Fitzpatrick function of T is [4] ϕ T R X X, ϕ T (x,x ) := sup x y,y x + x,x (2) (y,y ) T and the Fitzpatrick family associated with T is F T := h R h is convex and l.s.c. X X h(x,x ) x,x, (x,x ) X X (x,x ) T h(x,x ) = x,x In the next theorem we summarize the Fitzpatrick s results: Theorem 1.1 ([4, Theorem 3.10]). Let X be a real Banach space and T : X X be maximal monotone. Then for any h F T (x,x ) T h(x,x ) = x,x and ϕ T is the smallest element of the family F T. Fitzpatrick s results described above were rediscovered by Martínez-Legaz and Théra [9], and Burachik and Svaiter [2]. It seems interesting to study conditions under which a convex function h R X represents a maximal monotone operator, that is, h F T for some maximal monotone operator T. Our aim is to extend previous results on this direction. We will need some auxiliary results and additional notation for this aim. The Fenchel-Legendre conjugate of f R X is f R X, f (x ) := sup x,x f(x). x X Whenever necessary, we will identify X with its image under the canonical injection of X into X. Burachik and Svaiter proved that the family F T is invariant under the mapping. J : R X X R X X, J h(x,x ) := h (x,x). (3) This means that if T : X X is maximal monotone, then [2] J(F T ) F T. (4) In particular, for any h F T it holds that h π, Jh π, that is, h(x,x ) x,x, h (x,x) x,x, (x,x ) X X.

M. Marques Alves, B. F. Svaiter / Maximal Monotonicity, Conjugation and... 555 So, the above conditions are necessary for a convex function h on X X to represent a maximal monotone operator. Burachik and Svaiter proved that these conditions are also sufficient, in a reflexive Banach space, for h to represent a maximal monotone operator [3]: Theorem 1.2 ([3, Theorem 3.1]). Let h R X X be proper, convex, l.s.c. and If X is reflexive, then h(x,x ) x,x, h (x,x) x,x, (x,x ) X X. (5) is maximal monotone and h,jh F T. T := {(x,x ) X X h(x,x ) = x,x } Marques Alves and Svaiter generalized Theorem 1.2 to non-reflexive Banach spaces as follows: Theorem 1.3 ([5, Corollary 4.4 ]). If h R X X is convex and then h(x,x ) x,x, (x,x ) X X, h (x,x ) x,x, (x,x ) X X (6) T := {(x,x ) X X h (x,x) = x,x } is maximal monotone and Jh F T. Moreover, if h is l.s.c. then h F T. Condition (6) of Theorem 1.3 enforces the operator T to be of type (NI) [6] and is not necessary for maximal monotonicity of T in a non-reflexive Banach space. Note that the weaker condition (5) of Theorem 1.2 is still necessary in non-reflexive Banach spaces for the inclusion h F T, where T is a maximal monotone operator. The main result of this paper is another generalization of Theorem 1.2 to non-reflexive Banach spaces which uses condition (5) instead of (6). To obtain this generalization, we add a regularity assumption on the domain of h. If T : X X is maximal monotone, it is easy to prove that ϕ T is minimal in the family of all convex functions in X X which majorizes the duality product. So, it is natural to ask whether the converse also holds, that is: Is any minimal element of this family (convex functions which majorizes the duality product) a Fitzpatrick function of some maximal monotone operator? To give a partial answer to this question, Martínez-Legaz and Svaiter proved the following results, which we will use latter on: Theorem 1.4 ([8, Theorem 5]). Let H be the family of convex functions in X X which majorizes the duality product: { } H := h R X X h is proper, convex and h π. (7) The following statements hold true:

556 M. Marques Alves, B. F. Svaiter / Maximal Monotonicity, Conjugation and... 1. The family H is (downward) inductively ordered; 2. For any h H there exists a minimal h 0 H such that h h 0 ; 3. Any minimal element g of H is l.s.c. and satisfies Jg g. Note that item 2. is a direct consequence of item 1.. Combining item 3. with Theorem 1.2, Martínez-Legaz and Svaiter concluded that in a reflexive Banach space, any minimal element of H is the Fitzpatrick function of some maximal monotone operator [8, Theorem 5]. We will also present a partial extension of this result for non-reflexive Banach spaces. 2. Basic results and notation The weak-star topology of X will be denoted by ω and by s we denote the strong topology of X. A function h R X X is lower semicontinuous in the strong weakstar topology if E(h) is a closed subset of X X R in the s ω topology. The indicator function of V X is δ V, δ V (x) := 0, x V and δ V (x) :=, otherwise. The closed convex closure of f R X is defined by clconvf R X, clconvf(x) := inf{µ R (x,µ) clconve(f)} where for U X, clconvu is the closed convex hull (in the s topology) of U. The effective domain of a function f R X is D(f) := {x X f(x) < }, and f is proper if D(f) and f(x) > for all x X. If f is proper, convex and l.s.c., then f is proper. For h R X X, we also define Pr X D(h) := {x X x X (x,x ) D(h)}. Let T : X X be maximal monotone. In [2] Burachik and Svaiter defined and studied the biggest element of F T, namely, the S-function, S T F T defined by S T R X X, S T := sup h F T {h}, or, equivalently S T = clconv(π +δ T ). Recall that J(F T ) F T. Additionally [2] and, in a reflexive Banach space, Jϕ T = S T. J S T = ϕ T (8) In what follows we present the Attouch-Brezis s version of the Fenchel-Rockafellar duality theorem:

M. Marques Alves, B. F. Svaiter / Maximal Monotonicity, Conjugation and... 557 Theorem 2.1 ([1, Theorem 1.1]). Let Z be a Banach space and ϕ,ψ R Z be proper, convex and l.s.c. functions. If is a closed subspace of Z, then λ[d(ϕ) D(ψ)], (9) inf ϕ(z)+ψ(z) = max (z ) ψ ( z ). (10) z Z z Z ϕ Given X, Y Banach spaces, L(Y, X) denotes the set of continuous linear operators of Y into X. The range of A L(Y,X) is denoted by R(A) and the adjoint by A L(X,Y ): Ay,x = y,a x y Y, x X, where X, Y are the topological duals of X and Y, respectively. The next proposition is a particular case of Theorem 3 of [10]. For the sake of completeness, we give the proof in the Appendix A. Proposition 2.2. Let X, Y Banach spaces and A L(Y,X). For h R X X, proper convex and l.s.c., define f R Y Y f(y,y ) := inf x X h(ay,x )+δ {0} (y A x ). If λ[pr X D(h) R(A)], (11) is a closed subspace of X, then f (z,z) = min u X h (u,az)+δ {0} (z A u ). Martínez-LegazandSvaiter[7]defined, forh R X X and(x 0,x 0) X X, h (x0,x 0 ) R X X h (x0,x 0 ) (x,x ) := h(x+x 0,x +x 0) [ x,x 0 + x 0,x + x 0,x 0 ] = h(x+x 0,x +x 0) x+x 0,x +x 0 + x,x. (12) The operation h h (x0,x 0 ) preserves many properties of h, as convexity and lower semicontinuity. Moreover, one can easily prove the following Proposition: Proposition 2.3. Let h R X X. Then it holds that 1. h π h (x0,x 0 ) π, (x 0,x 0) X X ; 2. Jh (x0,x 0 ) = (Jh) (x0,x 0 ), (x 0,x 0) X X.

558 M. Marques Alves, B. F. Svaiter / Maximal Monotonicity, Conjugation and... 3. Main results In the next theorem we generalize Theorem 1.2 to non-reflexive Banach spaces under condition (5) instead of condition (6) used in Theorem 1.3. To obtain this generalization, we add a regularity assumption (14) on the domain of h. Theorem 3.1. Let h R X X be proper, convex and h(x,x ) x,x, h (x,x) x,x, (x,x ) X X. (13) If is a closed subspace of X, then λpr X D(h), (14) is maximal monotone and Jh F T. T := {(x,x ) X X h (x,x) = x,x } Proof. First, define h := clh and note that h is proper, convex, l.s.c., satisfies (13), (14) and J h = Jh. So, it suffices to prove the theorem for the case where h is l.s.c., and we assume it from now on in this proof. Monotonicity of T follows from Theorem 5 of [7]. Note that for any x X Therefore, T(x) is convex and ω -closed. T(x) = {x X h (x,x) x,x 0}. To prove maximality of T, take (x 0,x 0) X X such that x x 0,x x 0 0, (x,x ) T (15) and suppose x 0 / T(x 0 ). As T(x 0 ) is convex and ω -closed, using the geometric version of the Hahn-Banach theorem in X endowed with the ω topology we conclude that (even if T(x 0 ) is empty) there exists z 0 X such that z 0,x 0 < z 0,x, x T(x 0 ). (16) Let Y := span{x 0,z 0 }. Define A L(Y,X), Ay := y, y Y and the convex function f R Y Y, Using Proposition 2.2 we obtain f(y,y ) := inf x X h(ay,x )+δ {0} (y A x ). (17) f (y,y) = min x X h (x,ay)+δ {0} (y A x ). (18) Using (13), (17) and (18) it is easy to see that f(y,y ) y,y, f (y,y) y,y, (y,y ) Y Y. (19)

M. Marques Alves, B. F. Svaiter / Maximal Monotonicity, Conjugation and... 559 Define g := Jf. As Y is reflexive we have Jg = clf. Therefore, using (19) we also have g(y,y ) y,y, g (y,y) y,y, (y,y ) Y Y. (20) Now, using (20) and item 1. of Proposition 2.3 we obtain and g (x0,a x 0 ) (y,y )+ 1 2 y 2 + 1 2 y 2 y,y + 1 2 y 2 + 1 2 y 2 0, (y,y ) Y Y (21) (Jg) (x0,a x 0 ) (y,y )+ 1 2 y 2 + 1 2 y 2 y,y + 1 2 y 2 + 1 2 y 2 0, (y,y ) Y Y. (22) UsingTheorem2.1anditem2. ofproposition2.3weconcludethatthereexists( z, z ) Y Y such that inf g (x0,a x 0 ) (y,y )+ 1 2 y 2 + 1 2 y 2 +(Jg) (x0,a x 0 ) ( z, z )+ 1 2 z 2 + 1 2 z 2 = 0. (23) From (21), (22) and (23) we have inf (y,y ) Y Y g (x 0,A x 0 ) (y,y )+ 1 2 y 2 + 1 2 y 2 = 0. (24) As Y is reflexive, from (12), (24) we conclude that there exists ( y, y ) Y Y such that g( y +x 0, y +A x 0) y +x 0, y +A x 0 + y, y + 1 2 y 2 + 1 2 y 2 = 0. (25) Using (25) and the first inequality of (20) (and the definition of g) we have f ( y +A x 0, y +x 0 ) = y +x 0, y +A x 0 (26) and y, y + 1 2 y 2 + 1 2 y 2 = 0. (27) Using (18) we have that there exists w 0 X such that f ( y +A x 0, y +x 0 ) = h (w 0,A( y +x 0 )), y +A x 0 = A w 0. (28) So, combining (26) and (28) we have h (w 0,A( y +x 0 )) = y +x 0,A w 0 = A( y +x 0 ),w 0. In particular, w 0 T(A( y+x 0 )). As x 0 Y, we can use (15) and the second equality of (28) to conclude that A( y +x 0 ) x 0,w 0 x 0 = y,a (w 0 x 0) = y, y 0. (29)

560 M. Marques Alves, B. F. Svaiter / Maximal Monotonicity, Conjugation and... Using (27) and (29) we conclude that y = 0 and y = 0. Therefore, As z 0 Y, we have z 0 = Az 0 and so that is, w 0 T(x 0 ), A x 0 = A w 0. z 0,x 0 = Az 0,x 0 = z 0,A x 0 = z 0,A w 0 = Az 0,w 0 = z 0,w 0, z 0,x 0 = z 0,w 0, w 0 T(x 0 ) which contradicts (16). Therefore, (x 0,x 0) T and so T is maximal monotone and Jh F T. Observe that if h is convex, proper and l.s.c. in the strong weak-star topology, then J 2 h = h. Therefore, using this observation we have the following corollary of Theorem 3.1: Corollary 3.2. Let h R X X be proper, convex, l.s.c. in the strong weak-star topology and h(x,x ) x,x, h (x,x) x,x, (x,x ) X X. If is a closed subspace of X, then λpr X D(h), is maximal monotone and h,jh F T. T := {(x,x ) X X h(x,x ) = x,x } Proof. Using Theorem 3.1 we conclude that the set S := {(x,x ) X X h (x,x) = x,x } is maximal monotone. Take (x,x ) S. As π is Gateaux differentiable, h π and π(x,x ) = h(x,x ), we have (see Lemma 4.1 of [5]) Dπ(x,x ) Jh(x,x ), where Dπ stands for the Gateaux derivative of π. As Dπ(x,x ) = (x,x), we conclude that Jh(x,x )+J 2 h(x,x ) = (x,x ),(x,x). Substituting Jh(x,x ) by x,x in the above equation we conclude that J 2 h(x,x ) = x,x. Therefore, as J 2 h(x,x ) = h(x,x ), S T. To end the proof use the maximal monotonicity of S (Theorem 3.1) and the monotonicity of T (see Theorem 5 of [7]) to conclude that S = T.

M. Marques Alves, B. F. Svaiter / Maximal Monotonicity, Conjugation and... 561 It is natural to ask whether we can drop lower semicontinuity assumptions. In the contextofnon-reflexivebanachspaces,weshouldusethel.s.c.closureinthestrong weakstar topology. Unfortunately, as the duality product is not continuous in this topology, it is not clear whether the below implication holds: h π? cl s ω h π. Corollary 3.3. Let h R X X be proper, convex and If h(x,x ) x,x, h (x,x) x,x, (x,x ) X X. is a closed subspace of X, then λpr X D(h) cl s ω h F T, where cl s ω denotes the l.s.c. closure in the strong weak-star topology and T is the maximal monotone operator defined as in Theorem 3.1: In particular, cl s ω h π. T := {(x,x ) X X h (x,x) = x,x }. Proof. First use Theorem 3.1 to conclude that T is maximal monotone and Jh F T. In particular, S T Jh ϕ T. Therefore, Jϕ T J 2 h JS T. As JS T = ϕ T F T and Jϕ T F T, we conclude that cl s ω h = J 2 h F T. In the next corollary we give a partial answer for an open question proposed by Martínez-Legaz and Svaiter in [8], in the context of non-reflexive Banach spaces. Corollary 3.4. Let H be the family of convex functions on X X bounded below by the duality product, as defined in (7). If g is a minimal element of H and λpr X D(g) is a closed subspace of X, then there exists a maximal monotone operator T such that g = ϕ T, where ϕ T is the Fitzpatrick function of T. Proof. Using item 3. of Theorem 1.4 and Theorem 3.1 we have that is maximal monotone, Jg F T and T := {(x,x ) X X g (x,x) = x,x } T {(x,x ) X X g(x,x ) = x,x }. As g is convex and bounded below by the duality product, using Theorem 5 of [7], we conclude that the rightmost set on the above inclusion is monotone. Since T is maximal monotone, the above inclusion holds as an equality and, being l.s.c., g F T. To end the proof, note that g ϕ T H.

562 M. Marques Alves, B. F. Svaiter / Maximal Monotonicity, Conjugation and... A. Proof of Proposition 2.2 Proof of Proposition 2.2. Using the Fenchel-Young inequality we have, for any (y,y ),(z,z ) Y Y and x,u X, h(ay,x )+δ {0} (y A x )+h (u,az)+δ {0} (z A u ) Ay,u + Az,x. Taking the infimum over x,u X on the above inequality we get that is, f(y,y )+ inf u X h (u,az)+δ {0} (z A u ) y,z + z,y = (z,z),(y,y ), (z,z),(y,y ) f(y,y ) inf u X h (u,az)+δ {0} (z A u ). Now, taking the supremum over (y,y ) Y Y on the left hand side of the above inequality we obtain f (z,z) inf u X h (u,az)+δ {0} (z A u ). (30) For a fixed (z,z ) Y Y such that f (z,z) <, define ϕ,ψ R Y X Y X, ϕ(y,x,y,x ) := f (z,z) y,z z,y +A x +δ {0} (y )+h(x,x ), ψ(y,x,y,x ) := δ {0} (x Ay). Direct calculations yields λ[d(ϕ) D(ψ)] = Y λ[pr X D(h)] R(A)] Y X. (31) Using (11), (31) and Theorem 2.1 for ϕ and ψ, we conclude that there exists (y,x, y,x ) Y X Y X such that Now, notice that infϕ+ψ = ϕ (y,x,y,x ) ψ ( y, x, y, x ). (32) (ϕ+ψ)(y,x,y,x ) f (z,z)+f(y,a x ) (z,z),(y,a x ) 0. (33) Using (32) and (33) we get Direct calculations yields ϕ (y,x,y,x )+ψ ( y, x, y, x ) 0. (34) ψ ( y, x, y, x ) = sup y, y A x + z, y + w, x (y,z,w ) = δ {0} (y +A x )+δ {0} (y )+δ {0} (x ). (35)

M. Marques Alves, B. F. Svaiter / Maximal Monotonicity, Conjugation and... 563 Now, using (34) and (35) we conclude that Therefore, from (34) we have y = 0, x = 0 and y = A x. ϕ ( A x,x,0,0) ( = sup y,z A x + x,x + Az,w h(x,w ) ) f (z,z) (y,x,w ) = h (x,az)+δ {0} (z A x ) f (z,z) 0, that is, there exists x X such that Finally, using (30) we conclude the proof. References f (z,z) h (x,az)+δ {0} (z A x ). [1] H. Attouch, H. Brezis: Duality for the sum of convex functions in general Banach spaces, in: Aspects of Mathematics and its Applications, J. A. Barroso (ed.), North-Holland Math. Library 34, North-Holland, Amsterdam (1986) 125 133. [2] R. S. Burachik, B. F. Svaiter: Maximal monotone operators, convex functions and a special family of enlargements, Set-Valued Anal. 10(4) (2002) 297 316. [3] R. S. Burachik, B. F. Svaiter: Maximal monotonicity, conjugation and the duality product, Proc. Amer. Math. Soc. 131(8) (2003) 2379 2383. [4] S. Fitzpatrick: Representing monotone operators by convex functions, in: Functional Analysis and Optimization, Workshop / Miniconference (Canberra, 1988), Proc. Cent. Math. Anal. Aust. Natl. Univ. 20, Australian National University, Canberra (1988) 59 65. [5] M. Marques Alves, B. F. Svaiter: Brøndsted-Rockafellar property and maximality of monotone operators representable by convex functions in non-reflexive Banach spaces, J. Convex Analysis 15(4) (2008) 693 706. [6] M. Marques Alves, B. F. Svaiter: A new old class of maximal monotone operators, J. Convex Analysis 16(3&4) (2009) 881 890. [7] J.-E. Martínez-Legaz, B. F. Svaiter: Monotone operators representable by l.s.c. convex functions, Set-Valued Anal. 13(1) (2005) 21 46. [8] J.-E. Martínez-Legaz, B. F. Svaiter: Minimal convex functions bounded below by the duality product, Proc. Amer. Math. Soc. 136(3) (2008) 873 878. [9] J.-E. Martínez-Legaz, M. Théra: A convex representation of maximal monotone operators, J. Nonlinear Convex Anal. 2(2) (2001) 243 247. [10] S. Simons: Quadrivariate versions of the Attouch-Brezis theorem and strong representability, arxiv:0809.0325 (2008).